Standalone RL-MPQ (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the Conservative scenario — a quantized variant of mistralai/Mistral-7B-v0.1.

Field	Value
Base model	mistralai/Mistral-7B-v0.1
Scenario	Conservative
Avg bits / weight	3.8125
Compression vs FP16	4.1967×
WikiText-2 PPL	5.1877
Layers	32
Bit distribution	`{'2': 1, '3': 4, '4': 27}`
Format	Fake-quant FP16 + `rlmpq_policy.json`

Collection: RL-MPQ — Mistral 7B — all five scenarios for Mistral 7B.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/mistral-7b-rlmpq-conservative"

model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo)

Other Mistral 7B scenarios

Scenario	Avg bits	Compression	WikiText-2 PPL
Aggressive	2.8125	5.6889x	10.025
Balanced	3.3438	4.785x	5.3053
Extreme Survival	2.5625	6.2439x	46.2801
High Fidelity	4.4375	3.6056x	4.8901

Grouped archive (all scenarios in one repo): AvoCahDoe/mistral-7b-rlmpq

Method

Phase 3 — PPO agent assigns per-layer bit widths under the Conservative reward target.
Phase 4 — Policy replayed on real weights; WikiText-2 perplexity validates quality.
Export — Fake-quantized FP16 weights compatible with Hugging Face Transformers.

Files

File	Description
`config.json`	Llama architecture + RL-MPQ metadata
`model.safetensors`	Fake-quantized weights
`rlmpq_policy.json`	Per-layer bit-width policy
`rlmpq_metrics.json`	Validation & PPL summary

Citation

bibtex

@misc{rlmpq_mistral_7b_conservative_2026,
  title  = {RL-MPQ Conservative: Mistral 7B Mixed-Precision Quantization},
  author = {AvoCahDoe},
  year   = {2026},
  url    = {https://huggingface.co/AvoCahDoe/mistral-7b-rlmpq-conservative}
}